I/O Performance Optimization Research Based On VT-d Technology Under Nested Virtualization

With the development of Virtualization Technology, the demands of live migration, development, test, maintenance and security control to Virtual Machine Monitor makes the Nested Virtualization becomes necessary. The nested I/O virtualization's primary implementation was based on software emulate and the paravirtualization on Xen project. The nested virtual device's DMA access leads CPU to trap and emulate the DMA frequently, brings high overhead for system and makes the I/O performance poor.Aiming at the above problem, this thesis takes the basic principle and main implementation of nested virtualization and I/O virtualization as the research object, reducing the VMexit times while nested virtual device‘s DMA access as the research direction, device direct access the underlying physical memory as research thought, according to the VT-d and shadow page table technology, designs a hardware assisted Virtual VT-d scheme, focuses on the optimization of nested I/O performance. The main research was as following:1. The thesis virtualizes the physical VT-d device, encapsulates the pass-through device to be a virtual VT-d structure, makes the nested VMM could pass the device to its guest, and creates a virtual DMA remapping structure which convert the nested device's DMA access to the underlying physical address.2. To accelerate the DMA access, shadow VT-d page table is designed to remap the nested GPA(Guest's Physical Address) to underlying VMM's MPA(Machine Physical Address) directly, makes the nested device access the physical device directly. Physical VT-d page table loads the mapping from nested device's GPA to underlying VMM's MPA, without intervention of nested VMM dealing with memory access, which could reduce VMexit times and the frequency of CPU scheduling.3. Recursive method is adopted to implement the nested device assignment, no need hardware to participate. Hiding the device and using the PCI(BAR) Base Address Register translation table in underlying VMM makes the device assigned to nested VMM. Then the same method is used to hide and assign the device to nested virtual machine recursively for CPU access to corresponding memory-mapped I/O and port I/O.This solution allows nested virtual machine directly access the hardware device, is expected to achieve single virtualization VT-d I/O performance. Takes the network as example, compared with pure emulated I/O devices about TCP and UDP, using Intel 82576 network card to test the network transmission, could improve by more than 33% in I/O throughput, the average CPU utility reduce 20%.